PRESCCO: Efficient Prediction Intervals under a Right-Censored Covariate

Abstract

In clinical studies, a patient's outcome, e.g., a cognitive test score, is typical or atypical depending on how it compares with the outcomes of patients at a similar point in a neurodegenerative disease, measured by how far they are from a common disease event. A prediction interval based on the time to that event provides that comparison, but that time is right-censored for most patients. Yet no method had computed such a prediction interval in this right-censored covariate setting. We first adapt three conformal prediction methods to this setting, and show that the estimated half-length, the distance the interval runs either side of its center, varies from one study to the next, so the same outcome can be judged typical in one study and atypical in another. We then develop the PRESCCO method, whose estimator of the half-length is semiparametrically efficient and doubly robust, staying consistent when one of its two models is misspecified, while the method loses no coverage. In simulations its standard deviation is four to thirty times smaller than under any conformal prediction method, and in a Huntington disease study with 77.2% right-censoring, an outcome typical in one study is no longer atypical in another.